Speaker Recognition based on Mel-Frequency Cepstral Coefficients and Vector Quantization
Mangesh Balpande, Rutik Sansare, Tushar Padelkar, Vishant Shinde · 2021 IEEE Bombay Section Signature Conference (IBSSC) · 2021
Speaker recognition is the process of recognizing the voice sample of a particular speaker from the set of speaker voice samples stored in the system. It can be done by using the parametric representation of speech signals to extract the various features using various methods. In this paper, we propose a voice recognition method which is implemented using Mel Frequency Cepstral Coefficients (MFCC) for feature extraction from voice samples and speaker modelling using Vector Quantization (VQ). LBG algorithm is used for the effective codebook generation in VQ. The system starts by storing the voice samples into the database, after purification of these samples the important features are extracted using the MFFC and by using the LBG algorithm in the VQ the centroid of the voice samples is calculated to recognize the speaker. The experimental results indicate that an average identification accuracy of 85-90% is achieved and hence this is one of the best methods for speaker recognition.